The truth about AI trading after a year testing. I've been building and testing trading algos with AI almost every day for close to a year. Thousands of strategies and backtests, around hundreds of strategy families, multiple live systems. Here's what nobody selling you an "AI trading bot" will tell you. First, there are 7 different ways to use AI in trading: 1. An LLM trades on its own 2. An LLM analyzes the market and recommends trades 3. An LLM helps you build and test a strategy 4. An LLM sits inside a deterministic strategy as an extra filter layer 5. LLM simulated environments (a la MiroFish) where the aggregate is used as a signal 6. Classical ML on top of your strategy (the Lopez de Prado pipeline: random forest, XGBoost, LightGBM) 7. Training your own transformer on proprietary data These are completely different things. Mixing them up is why most "AI trading" takes are garbage. What does NOT work: ❌ Letting an LLM trade or recommend trades. I ran ensembles of LLM agents predicting 1h, 4h and 24h direction. The win rate differences were pure bias and regime artifacts. Last month's best agent was noise. ❌ Training your own transformer. Unless you have truly proprietary data at scale, you're paying GPU bills to overfit. ❌ Looking for the infinite money glitch on a single asset with a single method. ❌ "Claude, build me a profitable strategy." It grabs some data, overfits, tests 50 hypotheses and hands you survivorship bias with a smile. Add ML on top and you get look ahead bias, leaky labels and no real out of sample. I ran this experiment properly. Thousands of backtests, hundreds of strategy families, frontier models. Strategies that survived honest validation with Sharpe above 0.8: zero. The models aren't dumb. Markets are just efficient, and AI is the most convincing overfitting machine ever built. What DOES work: 🟩 LLM as the engineer, not the trader. Build deterministic strategies, formalize hypotheses, iterate fast, but you own the validation. 🟩 LLM as one filter layer inside a rule based system (including simulated environments as that filter). The strategy has to work without it too. 🟩 The Lopez de Prado pipeline on top of an already working strategy. Purged cross validation, walk forward, deflated Sharpe. And still treat it as experimental until it survives out of sample. 🟩 Deploy lots of paper trading agents and let live performance rank them over weeks. Time beats backtests. 🟩 Copy trading as a feature inside your strategy. Validate your entries against what proven wallets are actually doing. --> Reality check: on the whole Hyperliquid leaderboard, the number of traders you can actually mirror is maybe 15 to 30. The rest are HFT you physically can't copy. The pattern is simple. AI accelerates the process, you own the edge through validation. The moment AI owns the edge, you own the losses. That's exactly why we're building @Trigr_xyz. An AI native strategy engine where every part of your strategy (signal, filter, risk) is a node that you or the AI can inspect, test and rewrite. Trusted backtesting with deflated Sharpe, purged walk forward validation and explicit overfit warnings, because a backtest without them is marketing. You deploy your strategies as paper agents and watch them prove themselves in real conditions before risking a dollar. And you validate trades against what proven traders are doing, built in.

Most fun thing to do regarding papertrade is for those stuck in queue to use the funds they wanted to gamble away and max leverage long BTC to cook the whole thing and get revenge on the whales

Any strategy that you built in pinescript or on local can now easily be imported and tested into Trigr via MCP Use them and build on top of them with our fully composable cross-asset strategy builder and backtester If you've ever tried to ask Claude "Build me a trading strategy for bitcoin, make no mistakes" --> Just use Trigr and it will actually output a working strategy

It's fantastic that companies emerge with massive funding within this vertical It will bring credibility to ai + trading and it will push all teams to build better products for users There is a wide range of angles and products you can build so I'm excited to test out Catalyst and see how it compares with @Trigr_xyz

In the world of crypto, liquidity and execution are becoming a commodity The new wave of apps will sell performance like what fomo is doing I believe there will be a wide range of verticals within this performance market based on users risk profile - copy trading - social - automation - ai - curation - yield - gambling - exotic markets - prop firms Every app builder should position within one of these verticals and fight to improve users performances rather than extract more for them For instance, I know that @ProprXYZ is focused on helping its users earn more and being fully transparent and that's why they are emerging as a category winner We'll follow the same approach with @Trigr_xyz for automated trading strategies

If you are tired of trenching, just deploy in Trigr

Altcoins are pumping now but wait for the new wave of onchain companies I'm very bullish on Umia's model 👀

Sometimes an edge is weird In an efficient market where everyone is over optimizing, doing the unoptimized thing is actually the edge

Not bad numbers for paper agents on @Trigr_xyz

That's how you start building @ProprXYZ x @Trigr_xyz

A polymarket/kalshi prop firm with api access, who is building this?